search

vita-epfl / social-transmotion

[ICLR 2024] Official implementation of "Social-Transmotion: Promptable Human Trajectory Prediction" in PyTorch.
GNU Affero General Public License v3.0
34 stars 2 forks source link

About LearnedTrajandIDEncoding #4

Closed rashidch closed 4 months ago

rashidch commented 4 months ago

Hi,

thank you for your awesome research project.

Can you please provide an explanation for why Trajectory and Person ID encoding are used together?

Why do you add learned encoding to (0 to 128), x[:, :, :, 0:half*2:2] and person encoding to (1 to 128), x[:, :, :, 1:half*2:2]

image

yanggao2000 commented 4 months ago

Hi @rashidch

Thanks for your interest!

As you mentioned we used one encoder for both trajectory information and person ID. Since trajectory data and person ID are relatively compact data compared with other visual cues (e.g., pose keypoints), encoding them together is more efficient than encoding independently with two different modules.

Hope this brief explanation helps you understand!

rashidch commented 4 months ago

Hi @yanggao2000

Can you also release code and prepared data for ETH/UCY dataset and Pedestrians and Cyclists in Road Traffic dataset?

yanggao2000 commented 4 months ago

Hi @rashidch

Unfortunately, we currently don't have plans to release the processed data for those datasets. However, one can obtain them by referring to our code for preparing the JTA dataset (https://drive.google.com/drive/folders/1iIp2B5y85OKZ7DkW7gBkmulUxMWW5SkY) and adjusting it, as we used the same method to prepare those data.